AI Paper Atlas

Research map / Video / Video generation

Veda: Scalable Video Diffusion via Distilled Sparse Attention

Shihao Han, Hao Yang, Xiaofeng Mei, Xinting Hu, Yi Jiang, Xiaojuan Qi

ICML 2026 · Video › Video generation · approach: Efficient video diffusion models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use oracle full-attention tile mask instead of average-pooling mask at fixed 90% sparsity (ablation)sparse attention in video DiT · video generation qualityvisual qualityImproved
add distilled tile scoring (Veda mask) over average poolingsparse attention · tile recall vs full-attention oracle (Fig. 5)tile recall: 34.2% -> 66.4%Improved
use triplet (mean, max, min) tile statistics instead of average pooling (ablation)Veda score estimator · estimator training losstraining loss: 0.965 -> 0.912Improved
add mean to max/min statistics (ablation)Veda score estimator · estimator training losstraining loss: 0.982 -> 0.912Improved
use head-aware tiling instead of best static tiling [4,4,8] (ablation)Veda · human evaluation, Waver-benchMotion Quality / Overall Quality: +7.2% Motion Quality, +9.6% Overall QualityImproved
stop gradient from estimator into backbone features (ablation)Veda training · generation qualitygeneration qualityImproved
Veda 90% sparsity vs full attentionWaver-T2V-1B · human SBS, Waver-bench 1.0visual quality preference: 50% vs 49%No clear effect
Veda vs VSA at equal 95% sparsityWaver-T2V-1B · human SBS, Waver-bench 1.0visual quality preference: 76% vs 24%Improved
raise Veda sparsity from 90% to 95% (ablation)Wan2.1-T2V-1.3B · VBenchsubject consistency: 0.887 -> 0.790Hurt
raise Veda sparsity from 90% to 95% (ablation)Waver-T2V-1B · VBenchsubject consistency: 0.940 -> 0.934Mixed
Oracle mask at 90% sparsityWan2.1-T2V-1.3B · VBenchaesthetic quality: 0.629 -> 0.325Hurt
sparse attention with Veda (95% sparsity) vs dense FA3Waver-T2V-12B · 720P 241 frameslayer latency: 1576.5ms -> 309.1msImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy